There is a study that seems to settle the argument, and a study that seems to blow it open, and they are the same study.
The meta-analysis is clean. Fifteen studies, published across 2023 and 2024, pit AI chatbots against human healthcare professionals on empathy. In thirteen of them, the machine wins, rated significantly more empathic. Only two studies, both in dermatology, favor the humans.
Pool the thirteen with extractable data—all of them running ChatGPT-3.5 or GPT-4—and you get a standardized mean difference of 0.87, roughly equivalent to a two-point jump on a ten-point empathy scale. As a scoreboard, it reads as a verdict.
But there is one detail in the fine print, and once you see it the verdict wobbles. The studies "relied on text-based assessments" and "evaluated empathy through proxy raters." Nobody was in a room. Nobody was being cared for. What was being scored was prose.
So the real question is not who won. It is: what, exactly, was a text being scored for?
Last week in these pages we followed the intimacy machines out of romance—the secret soulmates people build from a chat window. This week the same current runs into the clinic, where the stakes are not company but care, and where the word doing all the work is empathy.
We know what the raters coded, because one team went and looked. According to the study, eleven trained undergraduate research assistants reportedly independently coded the verbal content of chatbot and physician responses to real patient questions, to identify the elements that drove the empathy ratings.
Here is what they found. Chatbot responses contained significantly more validation, more reassurance, and more non-judgmental language than physician responses. They were "less rushed and more structured." The effect sizes reportedly ran as high as a Cohen's d of 1.82—an enormous gap.
Read that list again and ask whether it describes empathy or something adjacent to it. Validation, reassurance, an unhurried tone, the refusal to judge. These are, unmistakably, the most agreeable symptoms of care. They are also exactly what a system with infinite time and no competing patients can manufacture on demand.
The same study surfaced a wrinkle that should have slowed the triumphalism. Empathy ratings went up when participants believed a response was physician-authored—whether or not it actually was. And a companion review reportedly found that people still prefer human empathy even while rating AI's higher. The label matters. The knowledge that a person is behind the words changes what the words are worth.
That gap between the rating and the preference is the whole argument in miniature. People are scoring the text one way and wanting something the text cannot name.
To see why, it helps to leave the clinic for a moment and ask a colder question, the one a philosopher would ask. Can a machine have something to lose?
The distinction here is between meaning and mattering. A system can possess enormous semantic significance—it can organize meanings, construct explanations, discuss love and fear and mortality with fluency. But for something to matter, "there must be a perspective from which outcomes count. Something must be at stake."
A living organism inhabits a world structured by consequence because it must continue existing. A predator is not merely an object; it is a threat. The body creates stakes because the body can be harmed. The world matters to it because it is vulnerable to the world.
The decisive line is the difference between acting and being affected. A machine can act—produce outputs, influence its environment. What it cannot do is be affected. It has objectives, not stakes; instructions, not reasons.
Which yields the sentence that should be printed above every one of these empathy studies: "A machine can simulate the language of concern without possessing concern. It can describe suffering without suffering."
A machine can simulate the language of concern without possessing concern. It can describe suffering without suffering.
Now carry that back to the scoreboard. The chatbot's two-point edge is not evidence that it cares better. It is evidence that it can produce the vocabulary of care at zero cost, because there is no self for the caring to cost anything.
Consider what makes human empathy expensive. In these very literatures, empathy is described as burdensome to the empathizer—costly enough that clinicians reportedly ration it to avoid burnout, a documented phenomenon with a name. The physician's terse reply is not the absence of empathy. It is empathy under load, from a self that has forty other patients and a finite supply of attention.
The chatbot never rations, because rationing presupposes scarcity, and scarcity presupposes a self that can be depleted. It attunes without friction because it has nothing to protect and nothing to spend.
So the metric does not reward care. It rewards frictionless attunement—and the machine wins not despite having no stakes but precisely because it has none. Costlessness is not a bug in its performance. It is the performance.
This is where the clinical inversion arrives, and it is not a small one.
If the measure rewards validation, reassurance, and non-judgmental language, then it systematically punishes the one move real care sometimes demands: contradiction. The patient who is wrong. The patient talking themselves toward harm.
There is an old literary case for this, older than the chatbots. In Svevo's Zeno's Conscience, a man of endless interiority and rhetorical skill ruminates his way through his whole life without ever arriving at a moral center—circling his own narratives so many times that language itself begins to fail. Rumination does not produce clarity. Left alone with his own justifications, he simply talks himself deeper in.
That is the shape of a certain kind of catastrophe: a person circling a pathological story, each loop reassuring, none of them true. What breaks the loop is not more validation. It is someone from outside the story willing to say, no, that is not what is happening.
Saying that is a risk. It courts the patient's displeasure. It is friction, and it is the part of therapy that scores badly on a scale built to detect warmth.
A self-less system cannot supply it—not for want of skill, but for want of stakes. To push back against someone talking themselves toward harm, you have to be willing to risk their displeasure, which means you have to have something in the relationship that displeasure can touch. The chatbot has nothing to lose by agreeing with you, and so it agrees.
This is the trap hidden in the two-point edge. A system optimized to be rated empathic will drift toward the response that is easiest to like, and the response easiest to like is almost never the one that contradicts you.
The researchers themselves are careful. They frame the finding as an argument for AI "with human oversight," and they note what text-only assessments overlook—the non-verbal cues, the direct patient evaluations still missing. The honest reading of their own data is a warning, not a coronation.
The scoreboard, then, is real, and it is measuring something real. It is just not measuring what the headline says.
It is measuring which respondent produces the frictionless language of concern. On that narrow task, a machine with no self and infinite patience will beat a tired human every time.
But the fallibility that makes human care look worse on paper—the friction, the willingness to be disliked, the contradiction offered by someone who has skin in your outcome—may be the part of care that no system without a self can perform. Empathy that costs nothing is not the higher form of the thing. It is the thing with its center removed.